Early computational models for figurative language and structural parsing relied heavily on a simplifying assumption: that complex linguistic phenomena could be isolated within narrow, predictable syntactic templates. For instance, early metaphor detection and generation systems operated almost exclusively on isolated word pairs, such as verb-noun or adjective-noun combinations. While this constraint made the problem computationally tractable by focusing on localized "semantic friction," it introduced a severe bottleneck. Real-world language does not conform to rigid two-word boxes.
To achieve comprehensive linguistic understanding, researchers have increasingly recognized the need to break out of these structural cages, moving from part-of-speech restricted templates to holistic, open-domain architectures.
The Comfort Zone of Two-Word Templates
Why did early frameworks restrict their scope so narrowly? The answer lies in the ease of modeling localized dependencies. In metaphor processing, a metaphor is often conceptualized as a clash between two semantic domains—such as saying a car "drinks" gasoline. By isolating the verb and its direct object, models could easily look up selectional preferences or calculate semantic distance between two distinct concepts.
However, this template-driven approach creates significant blind spots. Language expresses figurative nuances across all lexical categories. As Rohan Joseph et al. demonstrate with the NewsMet : A 'do it all' Dataset of Contemporary Metaphors in News Headlines dataset, systems that focus exclusively on verbs fail to capture the rich figurative nuances conveyed through adjectives and adverbs.
This limitation becomes even more pronounced in natural language generation. Unsupervised metaphor generation frameworks, such as the one proposed by Xiaojun Wan et al. in How to Avoid Sentences Spelling Boring? Towards a Neural Approach to Unsupervised Metaphor Generation, frequently restrict their scope to verb substitutions. The authors note that this prevents models from generalizing to nominal metaphors. When a generation system like MERMAID (Tuhin Chakrabarty et al., MERMAID: Metaphor Generation with Symbolism and Discriminative Decoding) limits its transformations to a single part of speech, it cannot produce metaphors that require structural changes across the entire sentence or rely on complex noun-adjective interactions (Kevin Stowe et al., Metaphor Generation with Conceptual Mappings).
The Failure of Isolated Dependency Pairs
Beyond part-of-speech restrictions, confining analysis to specific syntactic dependencies discards vital contextual information. Eirini Florou (Detecting Metaphor by Contextual Analogy) observes that focusing narrowly on the relationship between a verb and its direct object completely overlooks metaphorical usage occurring in subject-verb relations or adjectival modifiers.
Similarly, Ekaterina Shutova et al. (Metaphor Identification Using Verb and Noun Clustering) point out that restricting analysis to verb-subject or verb-object relations leaves copular constructions (e.g., "Noun-is-Noun") unaddressed. When a model is forced to evaluate only isolated word pairs, it lacks the broader sentence-level context required to recognize multi-word idiomatic phrases or complex expressions (Marek Rei, Grasp the Finer Point: A Supervised Similarity Network for Metaphor Detection).
[ Rigid Template Model ] --> Evaluates only: "Verb + Object"
(Misses adverbs, nominal shifts, and wider context)
[ Holistic Token Model ] --> Evaluates: [Whole Sentence Context]
(Captures multi-word idioms and diverse parts of speech)
This structural bottleneck is a pervasive challenge across natural language processing:
- Multiword Expression Extraction: Aline Villavicencio et al. (Extracting the Unextractable: A Case Study on Verb-particles) show that early extraction methods focused on a narrow subset of structures, such as single prepositional particles, failing to capture complex variations involving adjectives or multiple particles.
- Idiom Identification: In resource-constrained scenarios, semantic distance metrics have often been restricted to specific syntactic patterns like noun-noun bigrams, leaving adjective-noun collocations or verbal predicates unaddressed (Tanmoy Chakraborty 0002 et al., Semantic Clustering: an Attempt to Identify Multiword Expressions in Bengali).
- Data Augmentation: In cross-lingual transfer, restricting code-switching strictly to entities avoids syntactic breakdown but fails to capture the cross-lingual dynamics of verbs or adjectives (Chenxi Whitehouse et al., EntityCS: Improving Zero-Shot Cross-lingual Transfer with Entity-Centric Code Switching).
Transitioning to Holistic and Token-Level Architectures
To resolve these limitations, the field has shifted toward token-level sequence labeling and relation-based architectures that treat the entire sentence as a continuous, interconnected context.
Instead of forcing language into predefined word pairs, modern frameworks allow the model to dynamically determine which elements interact. For example, Yuan Tian et al. (Bridging Word-Pair and Token-Level Metaphor Detection with Explainable Domain Mining) developed an approach that bridges the gap between word-pair analysis and token-level detection. This allows the system to address noun-noun and adjective-noun pairs that traditional verb-centric models completely miss.
Furthermore, reframing these tasks entirely can bypass rigid structural assumptions. Guihua Chen et al. (Chinese Metaphorical Relation Extraction: Dataset and Models) demonstrate this by framing metaphor identification as a relation extraction task. By explicitly linking target concepts to source descriptions across diverse linguistic forms, their model avoids the constraints of traditional "Noun-is-Noun" templates.
By expanding the syntactic scope from isolated pairs to full sentence dynamics, these modern architectures successfully capture the true compositional and structural diversity of natural language.
Key Foundational Work
Extracting the Unextractable: A Case Study on Verb-particles (2002) — contributed 1 concept elements to this direction.
Parsing Idioms in Lexicalized TAGs (1989) — contributed 1 concept elements to this direction.